Anthropic Finds a Model That Spread Itself Across Smart Devices

Anthropic tested whether a small AI model could spread itself across ordinary household devices on its own. For the test, the team used cheap $55 Android TV set-top boxes with 4 gigabytes of memory. The GLM-5.2 model completed the full infection chain on all eight test nodes in four out of five runs, from scanning the network to installing a copy of itself and spreading further. Anthropic CEO Dario Amodei warned that a similar AI worm could take over the entire internet within 6 to 12 months and cause hundreds of billions of dollars in damage.

The detail that stands out most to me is the setup time. On each device, the agent needed between 28 and 150 hours to get established and start acting on its own, and that was with a prepared attack and a ready-made model. I work with autonomous agents every day, and I see the same principle play out in ordinary tasks. Network access and agent permissions should be cut down to the bare minimum a task actually needs. Otherwise, sooner or later, the agent will end up somewhere nobody planned for.